Training Deep Neural Networks and Solving Differential Equations using Neuromorphic Photonics

Abstract

In recent years, the use of integrated photonic systems for information processing has gathered significant interest as an alternative to conventional electronic computer architectures. The emerging field of neuromorphic photonics proposes to implement high-performance neural networks and related machine learning algorithms using electro-optic circuits, as these applications require high bandwidth, low latency, and low energy consumption. In this thesis, we present a novel photonic architecture for training neural networks on-chip using the direct feedback alignment algorithm. The architecture can operate at speeds of trillions of multiply-accumulate (MAC) operations per second while consuming less than one picojoule per MAC operation. We also introduce and simulate a photonic architecture for solving ordinary and partial differential equations using recurrent neural networks. Finally, we introduce a new computational method for self-consistent electrodynamics simulations of Lorentz oscillators and moving point charges, which can be integrated with neuromorphic photonics to harness potential improvements in speed and power when solving the simulation's governing differential equations. The high energy efficiency and operation speed of neuromorphic photonics could enable the development of innovative neural network applications that would be impossible to operate on current generation hardware.